I’m not saying Elon is right here, but I’ve noticed that you seem to base a lot of your comments on what our current version of Autopilot can handle/understand. It’s been shown both in the talk Andrej gave in Feb 2020 and in his CVPR video that what they’re working on internally and what we have are two very different software branches. You can see this in how the birds-eye NN predictor sees the world and how much more accurate it is compared to how our current sees the world (falls apart when it tries to make sense of an entire scene like an intersection).
Not only that, but based on the CVPR video it shows their birds-eye view predictor handling multi-lane intersections and predicting their layouts incredibly well. This was not show in the Feb 2020 talk, so it seems they’ve made significant progress from Feb to June and it only stands to reason that they’re making good progress on other parts of FSD as well since they clearly aren’t standing still. Coupled with Elon’s recent comments about releasing the new Autopilot software stack in 2-4 months with “a lot of new functionality” and this just further reiterates that what we’re running is all based on an old stack that they are barely even working on anymore.
I know you’ve watched all these videos and understand that these things are being worked on, but it seems like you still think Tesla will continue to stick with their L2 Autopilot system. What makes you think that despite the progress they seem to be making?
Yes, I've seen the CVPR videos (I started that thread). I am aware of the NN work that Tesla is doing. And the work looks really good actually. But it's all perception stuff. Tesla does not seem to be doing a lot of work with planning and driving policy which is a huge part of autonomous driving.
For example, the NN prediction of intersections just deals with the problem of "seeing" an intersection, it does not really deal with the problem of handling the traffic around the car. For example, can the car safely handle an unprotected left turn where a cyclist cuts in front while you are half way into the intersection? Can the car handle an intersection that is under construction where the car needs to temporarily move over into the incoming traffic lane but there is a large truck partially blocking the view? These are the kind of driving scenarios that require more than just seeing an intersection, the car also needs that planning and driving policy to figure out the safest to way to handle what other vehicles are doing.
So again, it's awesome to see the cool perception stuff that Tesla is doing but we have no way of knowing how good the features will actually work in reality, especially in scenarios like the ones I mentioned. I don't think we can make "forward looking statements" and say that just because Tesla has NN that can predict intersections like 99% of the time, that out cars will be L5.
No, I do not think that Tesla will stay at L2 forever. I have no doubt that the rewrite and these new NN that we got a glimpse of at CVPR will add new functionality. Absolutely! In fact, I think we will get "turning at intersections" at some point. Eventually, Tesla will get to L3+. But getting to L5 is a huge task because of all the driving cases that need to be solved.
The reason I bring up current AP is because that is something I can actually judge. It does not require guesswork or forward looking statements. I can test how AP handles specific driving scenarios. And certainly, when Tesla releases new features I will eagerly test those. and if AP is able to handle an unprotected left turn with a cyclist that cuts in at the last second without driver intervention, I will be the first to applaud Tesla. I want to judge Tesla based on what they actually deliver to my car, rather than NN work that may or may not translate into a good feature.
Bottom line: L5 means being able to handle all roads, all driving cases, day and night, with no human intervention. That's a tall order. Tesla is not there yet. I am not saying Tesla won't make progress or even that Tesla won't achieve L4 or L5 someday. I am just saying let's not jump the gun. Being able to predict intersections is cool but there is a lot more that goes into L5.